Microeconomic Inventory Adjustment: Evidence From U.S.

Firm-Level Datay
Jonathan McCarthy Research Department Federal Reserve Bank of New York 33 Liberty Street New York, NY 10045 e-mail: jonathan.mccarthy@ny.frb.org Egon Zakrajek s Division of Monetary A airs, MS-84 Board of Governors of the Federal Reserve System 20th Street & Constitution Avenue, NW Washington D.C., 20551 e-mail: egon.zakrajsek@frb.gov February 2000

This is a substantially revised version of an earlier paper entitled Microeconomic Inventory Adjustment and Aggregate Dynamics." We thank Douglas Dwyer, Mark Gertler, Jim Kahn, and Ken Kuttner for helpful comments. Benjamin Bolitzer provided expert research assistance. All remaining errors and omissions are our own responsibility. The opinions expressed in this paper do not necessarily re ect views of the Federal Reserve Bank of New York, Board of Governors, nor the Federal Reserve System.
y

Abstract
We examine inventory adjustment in the U.S. manufacturing sector using quarterly rmlevel data over the period 1978 97. Our evidence indicates that the inventory investment process is nonlinear and asymmetric, results consistent with a nonconvex adjustment cost structure. The inventory adjustment process di ers over the business cycle: for a given level of excess inventories, rms disinvest more in recessions than they do in expansions. The inventory adjustment process has changed little between the 1980s and 1990s, suggesting that recent advances in inventory control have had little e ect on adjustment costs. Nevertheless, the optimal inventory-sales ratio in the durable goods sector has declined signi cantly during our sample period.
JEL Classi cation: D24, E22, E37 Keywords: inventories, S; s inventory policies, linear-quadratic model, business cycles

1 Introduction
The cyclical behavior of inventories has been an important feature of aggregate uctuations in the U.S. economy. Over fty years ago, Abramowitz 1950 showed that a plunge in inventory investment accounts for most of a contraction in output during a typical U.S. recession prior to World War II|a statistical nding that has continued to hold in the postwar data see, for example, Blinder and Maccini 1991. This empirical regularity of aggregate output uctuations has led macroeconomists to examine inventory investment as a potentially important channel for the propagation and the ampli cation of exogenous shocks to the economy. The vast majority of macroeconomic research concerned with inventory dynamics has utilized the linear quadratic L-Q model of Holt, Modigliani, Muth, and Simon 1960. In the canonical L-Q model, the convex production technology at the microeconomic level, combined with the representative agent assumption, yields aggregate inventory dynamics that have proved di cult to match with the data. In particular, as documented by Ramey and West 1999, the strong procyclical uctuations in inventory investment and the persistence of inventory movements conditional on sales are two features of the data that are di cult to reconcile with the L-Q model. In view of the L-Q model's poor empirical performance, a number of authors have argued that the model's lack of empirical success stems from the assumption of convex production technology at the micro-level; see, for instance, Blinder 1981, Caplin 1985, and Blinder and Maccini 1991. The linear inventory decision rules implied by this assumption, although relatively easy to aggregate and estimate, do not capture potential nonlinear features of microeconomic inventory behavior. Among others, these features include S; s type inventory policies, resulting from the presence of xed or proportional costs in the production technology, and asymmetries or irreversibilities induced by di ering costs between drawing down and expanding inventory levels.1 Our goal in this paper is to examine the extent to which rm-level inventory adjustment in the U.S. manufacturing sector di ers from the adjustment process implied by the L-Q model. We embed an inventory problem into a exible and empirically tractable framework|the so-called generalized S; s approach|developed in a series of papers by Caballero, Engel, and Haltiwanger.2 A fundamental element of this approach is the adjustThe possibility of technological nonconvexities in the context of the L-Q model, in particular increasing returns to scale, is examined by Ramey 1991, who nds industry-level evidence consistent with rms operating in a region of declining marginal costs. 2 For a general theoretical treatment of adjustment and aggregate dynamics in S; s economies, see Caballero and Engel 1991. For applications of the generalized S; s approach to business xed investment, see Caballero 1999, Caballero and Engel 1999, and Caballero, Engel, and Haltiwanger 1995; employment dynamics are examined by Caballero and Engel 1993 and Caballero, Engel, and Haltiwanger 1997.
1

1

ment function of the control variable in question. In our model, the inventory adjustment function relates the fraction of the deviation between the target" and the actual level of inventories that rms close during a period to the size of that deviation. The main advantage of our approach is that the inventory adjustment function is estimated nonparametrically and, therefore, can take on a wide variety of shapes, including those implied by the L-Q model, a simple S; s model, and a more general nonlinear, asymmetric adjustment model. Our analysis of inventory investment dynamics applies the generalized S; s framework to a long panel of high-frequency rm-level data for the U.S. manufacturing sector. Compared to other micro-level studies, our rm-level data set has several advantages for studying aggregate inventory behavior. First, our data set comprises a large portion of the U.S. manufacturing sector, averaging over 60 percent of aggregate manufacturing inventories. Second, our data cover the time period from 1978 to 1997 and thus include a number of business cycles. Third, the frequency of observation is quarterly, which is of greatest relevance for studying aggregate business cycle uctuations. Contrary to the predictions of the L-Q model, our results indicate that signi cant nonlinearities and asymmetries exist in the estimated inventory adjustment functions. The nonlinearity is consistent with the use of S; s type inventory policies, while the asymmetries suggest the presence of irreversibilities in the production technology. Given the prominence of inventories in aggregate uctuations, we examine the adjustment function at various stages of a business cycle. We nd that for a given level of excess inventories, rms reduce their inventory holdings more in recessions than they do in expansions, a result consistent with the cyclical pattern of inventory investment in the aggregate data. The long time span of our panel also enables us to examine the extent to which the inventory adjustment process may have changed over time. In particular, there has been considerable debate whether improvements in inventory control methods during the 1980s e.g., just-in-time techniques, bar coding, etc. may have muted the inventory cycle, translating into reduced volatility of aggregate output uctuations; see, for instance, Filardo 1995, Allen 1995, and McConnell and Perez-Quiros 1998. Our results paint a mixed picture. Although the rates of adjustment toward the target inventory levels are somewhat higher in the 1990s than in the 1980s, the decade-speci c adjustment functions exhibit very similar shapes, indicating that the inventory adjustment process has not changed signi cantly during our sample period. Moreover, the cross-sectional dispersion of inventory deviations from their target levels has not declined over time, a result consistent with a stable inventory adjustment process. However, the average target inventory-sales ratio in the durable goods sector has declined signi cantly during our sample period, indicating that advances in inventory control have had a sizable e ect on this margin of rms' inventory behavior. The relatively stable adjustment cost structure on the 2

one hand, and the declining level of target inventory holdings relative to sales on the other, together imply that the e ect of inventory control improvements on aggregate inventory uctuations remains an open question. The rest of the paper is organized as follows. In the next section, we discuss the issues raised by linear versus nonlinear inventory control. In Section 3, we outline our approach to the study of inventory dynamics and discuss the details behind the construction of our key state variable|the deviation between actual and target levels of inventories. In Section 4, we provide a brief description of the data and econometric issues. In Section 5, we present our main results, and Section 6 concludes.

2 Linear versus Nonlinear Inventory Control
As discussed in the introduction, a majority of empirical inventory research is based on the L-Q model. The key assumption of this model is that rms maximize pro ts subject to a convex production technology. This assumption implies that rms will attempt to smooth production in the face of stochastic sales. Embedding the maximization problem into a representative agent framework and using quadratic functional forms to approximate the cost structure associated with the convex production technology, the L-Q model yields a microstructure that is relatively easy to aggregate and estimate. Although this microstructure has provided substantial insights into rm and aggregate inventory behavior, it has had di culty explaining two fundamental and robust features of inventory investment. First, inventory investment is highly procyclical: inventories tend to be built up gradually in expansions and to be drawn down rapidly in recessions. Second, inventory movements exhibit considerable persistence, even after conditioning on sales.3 Within the linear-quadratic representative-agent framework, two explanations have been advanced to reconcile these two facts with the model. The rst explanation assumes that highly persistent exogenous shocks a ect the cost of production. These cost shocks cause rms to bunch production|leading to procyclical inventory investment|while the built-in persistence of the cost shock process is transmitted to the inventory-sales relationship. The second explanation posits a strong accelerator motive and high costs of adjusting production. The accelerator motive, which links current inventories to expected future sales, and positively serially correlated sales cause inventory investment to be procyclical. The high adjustment costs imply that a return to the long-run equilibrium following a
Despite mixed evidence that inventories and sales are cointegrated at the industry or the aggregate level, the presumably stationary linear combinations of inventories and sales|the so-called inventory-sales relationship in the Ramey and West 1999 terminology|exhibit very high rst- and second-order autocorrelations, even at an annual frequency, indicating that the adjustment to long-run equilibrium takes place over many periods.
3

3

perturbation of the inventory-sales relationship will be gradual, because rms will adjust production slowly to minimize costs. Although each explanation is economically plausible, the empirical support for either is tenuous at best. The persistent cost shock hypothesis seems to work only when these shocks are modeled as unobservable to the econometrician. Although it is plausible that the rm's cost structure could be a ected by unobservable disturbances, their observable counterparts such as real unit labor costs and interest rates appear to have no appreciable e ect on inventory investment. Evidence in favor of the adjustment cost hypothesis is equally unpersuasive. Estimates of adjustment cost parameters are unstable across different speci cations and estimation techniques and range from negligible to economically implausible.4 In contrast to the L-Q vein of inventory research, a considerable operations research literature, starting with Scarf 1960, emphasizes xed costs and other nonconvexities in the production planning problem.5 From the macroeconomic perspective, Blinder 1981 and Blinder and Maccini 1991 have provided compelling arguments that nonconvexities in the production process may be crucial for our understanding of aggregate inventory dynamics. The optimal inventory policy in such an environment is of the S; s type, implying periods of inaction when inventories are depleted, followed by periods of activity during which inventories are replenished. Such nonlinear microeconomic inventory behavior has the potential to a ect aggregate inventory dynamics and, furthermore, calls into question the representative agent assumption underlying most applied macro-inventory research. If rms follow state-dependent S; s policies, the time-path of the entire cross-sectional distribution of inventories has an e ect on aggregate dynamics. In such an economy, a negative aggregate shock could result in fewer rms reaching their trigger inventory levels, exacerbating the decline in inventory investment beyond what would be expected under a representative agent L-Q model. In part because of these complications, empirical macroeconomists have been reluctant to forsake the analytically tractable L-Q framework for S; s type models. Although most economists would agree that a rm's production inventory problem will likely di er from
As is the case in all applied work, both at the micro- and macro-level, the inventory literature is not immune to serious measurement problems in the data that undoubtedly contribute to the poor empirical performance of the L-Q model. Some studies try to mitigate this problem by using the presumably more accurate, though considerably more limited, physical product data. These studies nd somewhat greater support for the L-Q model, in particular for the production smoothing motive; see Fair 1989 and Krane and Braun 1991 for examples of this approach. Alternatively, Schuh 1996 dispenses with the representative agent assumption. Using monthly plantlevel data, Schuh 1996 estimates the L-Q model and nds that accounting for the aggregation bias|in both the cross-sectional and the time-series dimensions|moderately improves the t of the L-Q model. 5 See Hordijk and Van der Duyn Schouten 1986 for a general treatment of continuous review inventory models with xed costs.
4

4

the assumptions underlying the L-Q model, they argue that the aggregation across goods and time may smooth the e ects of any nonconvexities. As a result, the L-Q model remains a reasonable approximation to the inventory behavior of interest to macroeconomists; see, for example, Blanchard 1983 and Ramey and West 1999. Moreover, at the macro-level, empirical evidence supportive of S; s type models is limited.6 In particular, using the theoretical aggregation results for dynamic S; s economies derived by Caplin 1985 and Caballero and Engel 1991, several studies have attempted to test the predictions of a simple S; s model using both aggregate and rm-level data for the trade sector.7 The basic conclusion of these studies is that the time-series behavior of inventory investment in the trade sector is consistent with the S; s model's steady-state implications. The problem with this conclusion, however, is twofold. First, the aggregate steady-state results of Caplin 1985 are valid only under restrictive assumptions of exogenous, serially uncorrelated sales, time-invariant S; s bands, and no delivery lags.8 Second, the steady-state, reduced-form relationships between inventory investment and sales are also consistent with an economically plausible parameterization of the L-Q model with stockout costs; see, for example, Blinder 1986, Kahn 1987, and Krane 1994.9 Therefore, if we are to shed light on whether nonconvexities matter for inventory behavior at business cycle frequencies, we must study inventory dynamics rather than steady-state behavior. Because structural S; s type models must be kept simple to derive analytic results useful for empirical analysis, the data almost surely would reject such models. Thus, in our approach, we will sacri ce some structural rigor to provide a tractable empirical framework, which nonetheless encompasses the possibility of nonlinear behavior.

3 Inventories and the Generalized S; Approach
s

In this section, we describe the generalized S; s approach used to analyze rm-level inventory investment decisions. The rst subsection discusses the basic elements of our model, while the second discusses the speci cs in measuring the state variable of the model.
6 At the micro-level, Hall and Rust 1999 examine nearly two years of daily transactions data from a steel wholesaler. They nd the rm's inventory order policies are consistent with a state-dependent S; s model. However, their data are probably too limited in scope to change the opinions of those skeptical about using these models to study aggregate inventory behavior. 7 See Mosser 1988, 1991, Episcopos 1996, and McCarthy and Zakrajek 1997. s 8 Mosser 1988, 1991 and McCarthy and Zakrajek 1997 attempt to allow for delivery lags and serial s correlation in sales and continue to nd support for the S; s model. 9 To address this concern, McCarthy and Zakrajek 1997 also estimate an Euler equation associated s with the L-Q model. They nd that, although the model is rejected at the industry level, estimates of the structural cost parameters are economically reasonable and statistically signi cant at the rm level; however, the over-identifying restrictions imposed by the model are rejected, and the parameter estimates are not stable across di erent asymptotically equivalent normalizations.

5

3.1 Basic Elements
The premise underlying our analysis is that a rm's inventory adjustment within a period depends upon the size of its perceived inventory shortfall or excess. In particular, we allow for the possibility that rms may respond to a greater degree as inventories move further away from their target levels. If such nonlinear inventory adjustment is detected, it would be indicative of nonconvexities in the production process. At the microeconomic level, inventory investment in our model depends upon a single state variable|the log deviation between target and actual levels of inventories, which we label as the inventory deviation index z: 
zit lnHit , lnHit,1 : 

1 

In equation 1, i indexes rms, t indexes time, and Hit and Hit denote actual and target real end of period t inventory stocks, respectively. It is important to note that because zit  depends on Hit , a theoretical construct, it is model-dependent. Although rms may continuously monitor their inventories and sales, observations in the model are at discrete intervals. Therefore, the timing of shocks and adjustment within a period needs to be spelled out. In our timing conventions, we follow Caballero, Engel, and Haltiwanger 1997 and assume that although rms may experience a sequence of shocks within a period, these shocks can be summarized by a single aggregate shock t common to all rms and a single idiosyncratic shock it that has a zero cross-sectional mean in each  period. After observing both shocks, rms determine their new target inventory levels Hit and thus their inventory deviation zit . Firms then adjust their inventory levels accordingly, and the process is repeated. The evolution of zit over time thus re ects the shocks to a rm's target inventory level and its adjustments in response to these shocks. Using equation 1 and our within-period timing assumptions, we can decompose the change in a rm's inventory deviation index between periods t , 1 and t, zit , as  zit = lnHit , lnHit,1 =  t + it  , lnHit,1 : 

2

The next element underlying our analysis is the cross-section of rms' inventory deviations in period t, denoted by fz; t. This is the cross-sectional probability density of rms' inventory deviations immediately preceding inventory adjustments during period t. Therefore, the fraction of rms with inventory deviations between z and z + dz in period t is approximately equal to fz; tdz. The last element in our framework is the inventory adjustment function, denoted by z; t. This function maps the inventory deviation z to the fraction of that deviation that 6

is closed by a rm within a period. To compute it, we determine which rms have an inventory deviation close to z in period t and calculate the fraction of the deviation closed on average by those rms. It then follows that the average inventory growth rate for rms with inventory deviation z in period t is equal to zz; t. Much of our empirical analysis will focus on the shape of the adjustment function z; t, because its shape provides considerable information about the nature of production technology. In particular, a constant adjustment function, z; t = 0 for all t, is consistent with a time-invariant convex adjustment technology and generates aggregate inventory dynamics identical to those implied by a representative-agent model with quadratic adjustment costs; see, for instance, Caballero and Engel 1993. Therefore, if the estimated adjustment function di ers from a constant function, this provides evidence consistent with the presence of nonconvexities in the inventory adjustment process. To conclude this subsection, we discuss how the three elements underlying our analysis combine to relate rm-level inventory investment decisions to aggregate inventory dynamics. Letting lnHtA denote the aggregate growth rate of inventories in period t, the preceding de nitions imply that Z A = zwz; tz; tfz; tdz; lnHt 3 where wz; t  0 for all t is a weighting function. Equation 3 shows that the dynamics of aggregate inventory investment are determined by the interaction between the adjustment function and the shifts in the cross-sectional density of deviations induced by aggregate and idiosyncratic shocks. In general, as long as the adjustment function z; t depends explicitly on z, aspects of the cross-sectional distribution other than its mean e.g., dispersion and skewness will in uence aggregate inventory dynamics.10

3.2 Measuring Inventory Deviations
A key to our analysis is the construction of a measure of the inventory deviation index z. To  do so, we need to formulate and estimate a model for Hit , the unobserved target inventory level. In formulating such a model, we assume that in the absence of adjustment costs, a rm's objective is to balance the costs of a potential stockout with the costs of holding inventories. As in the standard L-Q model, this tradeo can be captured by the following
In fact, Caballero and Engel 1993 and Caballero, Engel, and Haltiwanger 1995, 1997 test this implication for aggregate investment and employmentPp assuming a time-invariant, polynomial approximation by to the adjustment function|that is, z; t = j=0 j zj for all t. Under such assumption, equation 3 implies that the aggregate inventory growth depends linearly on the rst p + 1 weighted moments of the cross-sectional distribution fz; t; for a criticism of such regressions, see Veracierto 1998.
10

7

quadratic cost function: 1 Chit; mit ; sit+1  = 2 hit , Amit  + iEit sit+1 2 ; 4

where hit denotes the logarithm of inventories at the end of period t, sit+1 denotes the logarithm of real sales in period t+1, Eit is the expectations operator for rm i conditional on all the information through the end of period t, and i 0 is a rm-speci c cost elasticity with respect to expected future sales. The function A governs the signi cance of stockout costs relative to holding costs and depends on the vector of variables mit .11 The function A in equation 4 determines the optimal inventory-sales relationship. We allow this inventory-sales relationship to depend on the vector of rm-speci c variables mit . The source of this variation could re ect, for example, idiosyncratic and aggregate shocks to the cost of production. By a ecting the opportunity costs of holding inventories, such cost shocks may induce variation in stockout costs relative to holding costs.12 In addition, the optimal inventory-sales ratio is likely to vary because of seasonal factors. Finally, changes in inventory control techniques may induce long-run trends in the optimal inventory-sales relationship that are not captured by other variables. To allow for such variation in the inventory-sales relationship, we assume that mit = d0t ; cit ; cit+1 0 , where dt is a k-vector of nonstochastic variables to be speci ed later, cit is the logarithm of the cost per unit of output in period t, and we let A take on the following log-linear form:

Adt ; cit ; cit+1  = 0idt + i cit + i Eit cit+1 ; 

5

where i is the k-vector of unknown rm-speci c parameters corresponding to the vector of nonstochastic variables dt . Firm-speci c parameters i 0 and i 0 measure the extent to which inventories are used to bu er production from cost shocks. Using equations 4 and 5 in a standard intertemporal cost minimization problem, it follows that the optimal log-level of inventories h satis es the rst-order condition, it

h = 0i dt + i sit + i cit + i Eit sit+1 + iEit cit+1 : it 

6

Using equation 6, we can now derive our measure of the inventory deviation index z. First, let eit denote the inventory deviation at the end of period t, after the adjustment has taken
Compared to the typical accelerator term in the L-Q model, our speci cation of the tradeo between inventory holding costs and stockout avoidance is somewhat nonstandard, as all variables are stated in logarithms. Hence in our speci cation, the rm's cost function depends on the percentage di erence between the target and actual inventory levels rather than the absolute di erence between the two. With rm-level data, we believe that the log-speci cation provides a more natural description of these costs. 12 Speci cally, higher current costs would tend to reduce the current target level of inventories given sales, whereas greater expected future growth of costs would tend to increase the current target inventories given sales; Kahn 1992 derives a version of the L-Q model that formalizes this intuition.
11

8

place. It then follows that, or

eit = h , hit ; it 

7 8

eit = 0i dt + i sit + icit + i Eit sit+1 + iEit cit+1 , hit :

Note that from equations 1 and 7, eit di ers from zit only in that the former incorporates the inventory adjustment during the period|that is, eit = zit , hit . In our model, the time-series properties of the ex post inventory deviation eit are crucial for identifying the parameters of the model and thus the optimal inventory level. With continuous monitoring of inventories and observations at discrete intervals, the post-adjustment deviations re ect two factors. First, although we assume that all the adjustment occurs at the end of the period, the actual rm-level inventory investment decisions in our data may take place at times that do not coincide with the observation interval. Second, rms may not adjust inventories to their optimal level. In this context, we assume that deviations following the adjustment do not persist inde nitely and that neither factor causes a systematic bias in the observed ex post inventory deviations. This implies that for each rm i, eit is a realization from a stationary stochastic process, with an unconditional mean equal to zero, E eit = 0, and nite variance, E e2 = i2 , for all i. Therefore, eit can be considered a random disturbance in the following it regression: hit = i0 dt + isit + i cit + i Eit sit+1 + iEit cit+1 , eit : 9 To estimate equation 9, we need to specify the k-vector of nonstochastic variables dt and parameterize the exogenous forcing processes for sit and cit . We assume that the deterministic part of the stockout avoidance parameter, dt , consists of low- and highfrequency components. The low-frequency component includes linear and quadratic time trends and re ects such in uences as advances in inventory monitoring technology, changes in the rm's relationship with its suppliers, and changes in product diversity over time. The high-frequency component captures movements in the stockout avoidance behavior associated with seasonal uctuations. Thus, dt = qt0 ; t; t2 0 , where qt = q1 t; : : : ; q4 t0 is a vector of quarterly indicator variables. Finally, we assume that both the growth of sales, sit , and the growth of costs per unit of output, cit , can be represented by a stationary, rm-speci c AR4 process.13 Thus, the conditional expectations of sit+1 and cit+1 in equation 9 can be replaced by four lags of these variables. Making this substitution, we obtain the regression equation used to
We have experimented with the AR speci cations of lag length 2; 3; : : : ; 6 with a negligible e ect on our results.
13

9

estimate the inventory deviation index z,
X X hit = 0i dt + i sit + icit + i sit+1, + 'i cit+1, + uit ;

4

4

=1

=1 

10

where uit = ,eit , and the reduced-form parameters i and 'i , = 1; : : : ; 4, are a combination of structural parameters i and i and the parameters of the exogenous forcing processes for sit and cit , respectively. Note that in equation 10, the structural parameter i is not identi ed. The negative of the estimated residual uit from equation 10 is our estimate of eit . To ^ derive an estimate of zit , the state variable in our model, recall that equations 1 and 7 imply that zit di ers from eit only in that the latter incorporates adjustment. Therefore, using our estimate of the ex post deviation eit ,

zit = eit + hit; ^^
where hit denotes the growth rate of inventories of rm i in period t. 

11

4 Data and Econometric Methodology
4.1 Data
The data set used in our analysis come from the Standard and Poor's COMPUSTAT quarterly P S T, Full Coverage, and Research data les. The data set consists of a panel of 2,169 manufacturing rms and covers the time period 1978:Q1 to 1997:Q4 80 quarters. The panel is unbalanced, with the minimum continuous tenure of 20 quarters. After eliminating rms with gaps in the time-series dimension or implausible entries, we are left with a total of 92,163 rm quarter observations. During our sample period, the rms in the panel account, on average, for 61 percent of aggregate non-farm manufacturing inventories. The Data Appendix contains details on the exact sample selection procedure, industry composition, the construction of variables, as well as summary statistics. Let us make one additional point concerning the data. If nonconvexities are important in determining inventory investment decisions, plant-level data may be preferable to rmlevel data. Many decisions concerning inventory investment are made at the plant level. For example, plants within a single rm that produce di erent products may have inventory policies that depend on the demand for the particular products that each plant produces. Thus, to the extent that individual plants operate as independent entities within a multiplant rm, the distribution of shocks and inventory deviations across plants within a rm may a ect rm-level inventory dynamics. 10

Moreover, nonlinear inventory adjustment is likely to be more apparent at less aggregated levels. For example, total inventories of the rm studied by Hall and Rust 1999 exhibit considerably smoother behavior at the daily frequency than do the inventories for each individual product category. Thus, concentrating our analysis at the rm level may lose some important information concerning the nonlinear adjustment at the fundamental microeconomic level. Nevertheless, many aspects of inventory decisions are centralized within the rm. Finished goods inventories may go to a centralized distribution center, which enables the rm to make production and inventory decisions more e ciently. Furthermore, nancial conditions and capital market access are rm- rather than plant-level phenomena. Hence, the well-documented sensitivity of inventory investment to movements in internal funds or net worth is evidence of the role that nonconvexities may play at the rm level. Thus, much of the e ect that nonconvexities may have on inventory adjustment can be studied at the rm level. Moreover, it is interesting to examine how much aggregation to the rm level, as well as time aggregation, convexi es the nonconvexities at less aggregated levels.

4.2 Econometric Methodology
In this section, we discuss the details behind estimation of equation 10. For notational convenience, let Xit = d0t ; sit ; cit ; sit ; : : : ; sit,3 ; cit ; : : : ; cit,3 0 denote the vector of all explanatory variables in equation 10, and let i =  0i ; i ; i ; i1 ; : : : ; i4 ; 'i1 ; : : : ; 'i4 0 denote the corresponding vector of parameters. Observing N rms over Ti , i = 1; : : : ; N, periods, we can write equation 10 for rm i compactly as,

hi = Xi i + ui; i = 1; : : : ; N: 

12

Our primary concern in equation 12 is the problem of parameter heterogeneity across rms. It is widely recognized that parameter heterogeneity at the micro-level|in both cross-sectional and time-series dimensions|can have important consequences for estimation and inference. As shown by Mairesse and Griliches 1990 and Schuh 1996, for instance, evidence of slope heterogeneity from rm-level panels is pervasive. A parsimonious speci cation that incorporates such parameter heterogeneity is the random coe cients model RCM; see, for example, Hsiao 1996. In the RCM speci cation, all the parameters of the model are random variables and stochastic speci cations are introduced to capture parameter heterogeneity. The RCM approach to parameter heterogeneity reduces the number of parameters to be estimated considerably while still allowing the coe cients to di er randomly across cross-sectional units and or time. To allow for variation in i across rms, we follow standard RCM assumptions that i = 11

+i , where i is a random vector with mean equal to zero that is distributed independently of the exogenous regressors Xi and the disturbances ui .14 Using these assumptions and stacking observations for all N rms yields,

h = X + Z + u; 

13

where Z = diag Xi i=1;:::;N is a block-diagonal matrix with the matrix Xi on its i-th diagonal 0 0 element,  = 1 ; : : : ; N 0 is an unknown vector of random e ects that capture parameter heterogeneity across rms, and u is an unknown random error vector. The foregoing assumptions imply that

E  = 0 ; and Var  = G 0 ; u 0 u 0R

" 

" 

" 

"  

14

where G =  IN , the symbol denotes the Kronecker product, and IN is the identity matrix of dimension N.15 Consistent and e cient estimates of the unknown parameter vectors and  in equation 13 can be obtained via Generalized Least Squares GLS. Formally, the GLS estimators of ^ and , denoted by ^ and  , respectively, are obtained by minimizing h , X 0 V ,1 h , X ; 15 with respect to , where V = ZGZ 0 +R is the covariance matrix of h. Minimizing equation 15, of course, requires the knowledge of V and, therefore, the knowledge of G and R. Given b b reasonable estimates of G and R, which are denoted G and R, respectively, we can obtain the feasible GLS FGLS estimates of the parameter vectors and , given by

b b ^ = X 0 V ,1 X , X 0 V ,1 h; bb^ ^  = GZ 0 V ,1 "; 

16 17

where " = h , X ^ and , denotes the generalized inverse. Using equations 16 and 17, our ^ estimate of the ex post inventory deviation index e is then given by, ^ e = ,h , X ^ , Z : ^ 18

Because of the unbalanced nature of out data, we use the Minimum Variance Quadratic
14 This independence assumption seems reasonable, because any suspected systematic dependence can, in principle, be explicitly modeled and hence accounted for. 15 Note that in equation 14, we let the covariance matrix  be unrestricted. Hence, we allow for the possibility that the elements of the random vector i may be correlated.

12

^ Unbiased Estimators MIVQUE of G and R to compute ^ and  . Developed by Rao 1971, the non-iterative MIVQUE of variance components are an alternative to the computationally intensive ML estimators. The procedure requires no distributional assumptions other than the existence of the rst four moments. Basically, the MIVQUE of a linear combination of the unknown variance components in G and R are obtained by nding a symmetric matrix W, such that Var h0 Wh is minimized subject to the conditions that h0 Wh is an unbiased estimator of the linear combination of the variance components and is invariant to any translation of the parameter.16 2 To simplify computations, we also assume that Var uit = i2 = e for all i. Hence, PN 2 R = e In , where n = i=1 Ti is the total sample size of our unbalanced panel. Because this homoscedastic assumption is likely to be violated in practice, we compute the covariance matrix of the parameter vector ^ using the asymptotically consistent estimator described by Diggle, Liang, and Zeger 1995. This estimator is computed as follows: 

"N b b ^^ b b ^ = X 0 V ,1 X , X Xi0 Vi,1 "i "0i Vi,1 Xi X 0 V ,1 X , ; Var

i=1 

19

where matrices with the subscript i correspond to those of the i-th rm. Table 1 contains the FGLS estimates of the structural parameters and from equation 10 for the major industry groups in our sample. Note that across all industries, estimates of the both parameters are of the right sign: 0 and 0. Almost uniformly, the estimates are statistically signi cant. Only for SIC 270 Printing & Publishing is the estimate of not statistically di erent from zero at the usual signi cance level, although it is of the right sign. In addition, there is considerable variation in the estimates of and across industries. Estimates of range from the low of 0.435 for SIC 320 Stone, Clay & Glass Products to the high of 0.925 for SIC 250 Furniture & Fixtures, while estimates of range from -1.258 for SIC 340 Fabricated Metal Products to -0.225 for SIC 270 Printing & Publishing. Using the parameter estimates from Table 1, we then can calculate the time path of optimal inventories for each rm. Figure 1 presents the cross-sectional average of the ratio of optimal inventories to real quarterly sales for the nondurable and durable goods sectors
As discussed by Baltagi 1995, the MIVQUE require a priori values of the variance components in G and Consequently, the resulting estimators are minimum variance only if these a priori values coincide with the true values. Two priors for the matrix W are typically used in practice: the identity matrix, denoted by MIVQUE0, and the method of moments ANOVA estimators denoted by MIVQUEA. We use the ^ MIVQUE0 estimators of G and R to obtain the FGLS estimators ^ and  in equations 16 and 17. Note that if one iterates on the initial values of the variance components until convergence, the MIVQUE will converge to ML estimates under normality; see Swallow and Monahan 1984 and Baltagi and Chang 1994 for a detailed discussion and relative e ciency comparisons of the various variance components estimators with unbalanced panel design.
16

R.

13

over our sample period. For the most part, the behavior of this average ratio is quite similar to that of the sectoral inventory-sales ratios computed from the published aggregate data. The average optimal inventory-sales ratio in the durable goods sector declines fairly steadily throughout our sample period, a pattern very similar to that in the aggregate data. The correlation between the two series is almost 0.9. In contrast, after a decline in the early 1980s, the average ratio in the nondurable goods sector has been relatively constant. Similarly, the aggregate inventory-sales ratio in the nondurable goods sector has been relatively constant throughout our sample period; the correlation between the two series in this sector is about 0.4 after 1983.17 The steady decline in the optimal inventory-sales ratio for the durable goods rms suggest one way that the recent improvements in inventory control may have a ected rmlevel inventory behavior. The wide-spread adoption of these innovations during the 1980s presumably has enabled rms to control and monitor their inventories better than they had previously. If so, rms then face a smaller probability of encountering a stockout for a given inventory level and thus would want to carry fewer inventories relative to their sales. If these practices became su ciently widespread throughout the sector, we would observe a decline in the average optimal inventory-sales ratio.18

5 Results
In this section, we present our main results. We characterize the main elements of equation 3, namely, the inventory adjustment function z; t and the cross-sectional distribution of inventory deviations, fz; t. Of particular interest are the shape of the inventory adjustment function and its variation across sectors and time. In addition, we examine the time path of the cross-sectional distribution of the inventory deviation index z.

5.1 Microeconomic Inventory Adjustment
To compute the inventory adjustment function z; t, we rst discretize the state space. The inventory deviation index z takes values between -0.4 and 0.4, over an equally-spaced grid with intervals of size 0.01. This interval covers over 99 percent of the sample range
17 The aggregate inventory-sales ratio in the nondurable goods sector does not exhibit the pronounced decline in the early 1980s that is evident in the micro data; hence, the correlation over the entire sample period is only 0.1. However, this weak correspondence between the micro and aggregate data is a result of Standard & Poor's relatively limited coverage of nondurable goods rms early in the sample. In the mid 1980s, Standard & Poor's greatly expanded its coverage, especially among smaller nondurable goods rms, and the aggregates computed from the micro data are much more closely correlated with the aggregate time series. 18 Abernathy et al. 1999 contains a detailed case study of the apparel and textile industry, where even in the face of product proliferation, lean retailing practices have been propagated through to the manufacturers.

14

of z. In each interval, we construct the value of the adjustment function by dividing the average inventory growth by z for those rms that are at z just before inventory adjustments take place.19 In what follows, all depicted adjustment functions are smoothed by a cubic B-spline.

5.1.1 Manufacturing Inventory Adjustment
The solid line in Figure 2 shows the estimated inventory adjustment function for the U.S. manufacturing sector. The two dotted lines represent two standard deviation error bands.20 The dashed line represents the smoothed density function of inventory deviations across all rms and quarters. Three observations about Figure 2 stand out. First, the adjustment function in nonlinear: rms adjust inventories more as the magnitude of the deviation increases. This nonlinearity is irrespective of whether the deviation is an inventory shortage z 0 or an inventory overhang z 0. The di erence between the adjustment rates for rms with large and small deviation is on the order of 15 percent and appears to be signi cant. The shape of the estimated adjustment function is consistent with the existence of nonconvexities in the production technology that induce rms to adopt S; s type inventory policies. Second, the inventory adjustment function is asymmetric. For small- to moderate-sized inventory deviations, rms with inventories above their desired level z 0 adjust less than rms with similar-sized inventory shortages z 0. There are several possible explanations for this asymmetry. First, because of a strong stockout avoidance motive, rms may be more willing to carry extra inventories. Second, market irreversibilities could prevent rms from reducing their excess inventories. Third, rms may be reluctant to cut output, because they nd it costly not to employ their capital and labor. The nal point about Figure 2 concerns the level of the estimated adjustment rates. In contrast to the implied adjustment rate estimated from a canonical L-Q model, the adjustment rates in Figure 2 are economically plausible. According to Ramey and West 1999, typical estimates of the adjustment rate from the L-Q model on quarterly U.S. manufacIn calculating the values of the adjustment function, values of z close to zero|that is, between -0.02 and 0.02|are excluded, because the calculation involves dividing the average adjustment rate by z. Because of the unbalanced nature of our panel and the varying degree of precision regarding our key parameter estimates, the average adjustment rate in each z-interval is computed as a weighted average of inventory growth rates, where the weights are given by the reciprocal of the rm-speci c standard deviation of the estimated post-adjustment inventory deviation index eit . All the results, however, are nearly identical ^ if an unweighted average is used in computations. 20 The error bands are obtained via a nonparametric bootstrap method. Speci cally, from the original sample, we draw with replacement the estimated inventory deviations and the actual inventory adjustments i.e., inventory growth rates. For each of the 5,000 bootstrap samples, we compute the adjustment function as described in the text. We then compute the standard deviation of the estimated average adjustment for each point in our z-space. Finally, the resulting 2 standard deviation error bands are smoothed using the same procedure as in the case of the adjustment function.
19

15

turing data are in the 10-20 percent range, indicating large costs of adjusting production. In our model, the estimated adjustment rates vary between 60 and 80 percent per quarter, an economically plausible range given the relative size of inventories to quarterly sales for the rms in our sample.

5.1.2 Sectoral Inventory Adjustment
In this section, we investigate whether the inventory adjustment function di ers between the durable and nondurable goods sectors. Sectoral di erences in the inventory adjustment process may provide some insight into the nature of the nonlinearities present in the overall inventory adjustment function. Apparent from Figure 3 is that the inventory adjustment functions for durable and nondurable goods rms exhibit a similar shape. In both sectors, the adjustment functions contain nonlinearities associated with generalized S; s type inventory policies. However, there are some noticeable di erences between the two sector-speci c adjustment functions. First, most obviously, the adjustment rate for nondurable goods rms is greater than that for durable goods rms for all values of the inventory deviation index z the di erence ranges from 5 to 10 percent. This di erence may re ect that nondurable goods manufacturers are more willing or better able to close inventory deviations, possibly owing to di ering production technologies or market structures. Second, the adjustment function for durable goods rms displays a greater asymmetry than the adjustment function for nondurable goods rms. This di erence suggests that the issues of stockout avoidance and market or production irreversibilities may be more important for inventory behavior of durable goods rms than for their counterparts in the nondurable goods sector. Moreover, the combination of lower inventory adjustment rates and greater asymmetries in the durable goods sector indicates that nonconvexities in the production technology may be more prevalent in durable goods industries.21

5.1.3 Time Variation in Inventory Adjustment
In this section, we examine the variation of the inventory adjustment function across time. First, we compare the adjustment functions during expansions and recessions to investigate how the adjustment process may di er over the business cycle. We then compare the adjustment functions estimated over the 1980s and the 1990s to examine the extent to which technological improvements in inventory control during our sample period may have a ected the inventory adjustment process.
There is persuasive evidence that nonconvex production costs play an important role in the automobile industry; see Bresnahan and Ramey 1994 and Hall 1999.
21

16

NBER-dated expansions and recessions during our sample period are displayed in Figure 4. The implied pattern of inventory investment underlying these functions is in accord with the stylized facts regarding the role of inventories in aggregate uctuations. Most notably, the adjustment rate for small to moderate inventory overhangs z 0 during recessions is above the adjustment rate for inventory overhangs of similar size during expansions. Thus, for a given level of excess inventories, rms reduce their inventory holdings more in recessions than they do during normal economic times. Furthermore, compared to a period of normal economic activity, target inventories are likely to be revised down in recessions. This macroeconomic state dependence could be a result of increased demand uncertainty or adverse cost or productivity shocks during periods of economic downturn. Consequently in recessions, a greater fraction of rms may nd themselves holding excess inventories. The interaction between the adjustment function and the cross-sectional distribution of inventory deviations at the onset of a recession would lead to a period of rapid aggregate inventory disinvestment. the inventory adjustment function across di erent subperiods of our sample. This exercise enables us to investigate the potential e ect of technological improvements and innovations in inventory management practices on the inventory adjustment process. Figure 1 of Section 4.2 shows that the estimated optimal inventory-sales ratio in durable goods industries has declined signi cantly during our sample period. We argued that this steady decline is consistent with the adoption of more e cient inventory control methods. In addition to this e ect, the impact of inventory control innovations on the adjustment process is equally important in assessing the implications of these innovations for aggregate inventory cycles. Suppose that following the adoption of modern inventory management practices, adjustment rates have increased signi cantly and the adjustment function has become closer to a constant function. Such pattern would imply that the costs of adjusting inventories have declined and that nonconvexities in the cost structure have a lesser e ect on rm behavior, resulting in reduced volatility of inventory investment. Given that our data cover nearly two full decades, we compare the inventory adjustment function estimated over the 1980s with that estimated over the 1990s. In light of the above-documented cyclical shifts in the inventory adjustment process, we exclude recession quarters when we estimate the decade-speci c adjustment functions. By doing so, we eliminate any di erences in the inventory adjustment process that would result from business cycle-dependent shifts in the adjustment function. The inventory adjustment functions estimated over the 1980s and 1990s are plotted 17

Cyclical Shifts in the Adjustment Function The inventory adjustment functions for

Long-Term Shifts in the Adjustment Function Next, we examine the stability of

in Figure 5.22 The two functions are similar, as they both display the same nonlinear, asymmetric shape as the adjustment function estimated over the entire sample period. Nevertheless, the adjustment rates are uniformly higher during the 1990s than during the 1980s, with the largest di erences occurring for inventory shortages z 0. Note also that relative to the 1980s, the adjustment function during the 1990s displays somewhat greater asymmetry between inventory investment and disinvestment. The fact that adjustment rates are somewhat higher in the 1990s than in the 1980s is consistent with the presumed e ects of improved inventory control on the inventory adjustment process. However, the shape of the adjustment function in the 1990s is also consistent with the continuing presence of nonconvexities and asymmetries in the inventory adjustment process. Combined with the relatively small increase in the rate of adjustment, this suggests that advances in inventory control methods have had little e ect on the adjustment cost structure over the past two decades. On the other hand, these improvements have likely had an e ect of lowering the optimal inventory holdings, particularly in the durable goods sector. The implications of these results for aggregate inventory uctuations are unclear. On the one hand, the recent advances in inventory control practices have done much to improve a rm's ability to continuously and accurately monitor their inventory levels, thus reducing the possibility of a surprise" shortage. This better information, in turn, reduces the perceived probability and the expected costs of a stockout, thereby reducing the need to carry as much inventory relative to sales as before. On the other hand, the production technology is likely to be una ected by such inventory control innovations, and the adjustment process given the desired inventory level will remain stable over time. The decline in the optimal inventory-sales ratio implies that inventories will be lower than they would be otherwise. However, with little change apparent in the adjustment technology, it is likely that the uctuations about the lower inventory targets may be of a similar magnitude as before, depending, ultimately, upon the distribution of inventory deviations and exogenous shocks.

5.2 Cross-Sectional Evolution of Inventory Deviations and Shocks
The cross-sectional distribution of inventory deviations, fz; t, is determined by the interaction between the aggregate and idiosyncratic shocks a ecting rms' inventory targets and the inventory adjustment process. The average density|where the average is computed over all rm quarter observations|is displayed by the dashed line in Figure 2. In
22 We also estimated the decade-speci c adjustment functions for the durable and nondurable goods sectors separately. Other than increasing noise in the data, controlling for sectoral di erences has little substantive e ect on our results.

18

this section, we consider the time variation of this distribution and the evolution of the shocks a ecting inventory deviations.

5.2.1 Cross-Sectional Moments of Inventory Deviations
Figure 6 displays the time paths of the cross-sectional weighted mean of inventory deviations and aggregate inventory growth for the durable and nondurable goods sectors.23 In both sectors, the cross-sectional mean of the inventory deviation index z and aggregate inventory growth move closely together; the correlation between the two series is 0.77 in the durable goods sector and 0.61 in the nondurable goods sector. The high correlation between the two series indicates that our estimates of the inventory deviations are successful in capturing the basic features of aggregate inventory investment. Nevertheless, the correlation is far from perfect, and there are episodes|in particular, the period surrounding the economic slowdown during the 1980s|when the behavior of the two series di ers markedly. This suggests that aspects other than the mean of the cross-sectional distribution of inventory deviations may a ect aggregate inventory dynamics. Moreover, the use of representative agent assumption in modeling aggregate inventory investment is unlikely to capture these missing features. The time paths of the weighted second, third, and fourth moments of the crosssectional distribution are presented in Figure 7.24 Each moment displays signi cant temporal variation and does not appear to be highly correlated with the other moments. Over time, the cross-sectional distribution of inventory deviations exhibits both substantial skewness and kurtosis; interestingly, these higher moments are more volatile in nondurable goods industries. The time variation of these higher moments and the imperfect correlation between the mean of the cross-sectional distribution fz; t and aggregate inventory growth together imply that idiosyncratic shocks have an important e ect on aggregate dynamics, an issue we examine further in the next section. Finally, note that the dispersion of inventory deviations, as measured by the crosssectional standard deviation, does not exhibit a secular decline in either sector during our sample period. A common conjecture concerning recent improvements in inventory control methods is that they make it easier for rms to align inventories more closely to their optimal levels. Thus, if advances in inventory management techniques had a signi cant impact on the adjustment process, we should, on average, observe a decline in the cross-sectional dispersion of inventory deviations. The fact that this decline has not occurred is consistent with the stability of the adjustment function over our sample period and indicates that
The means and all other moments are weighted by last period's rm size, measured by the level of inventory stocks and have been seasonally adjusted with quarterly dummies. 24 Excess kurtosis is de ned as the di erence between 3, the kurtosis of the normal distribution, and the sample kurtosis.
23

19

inventory control innovations have had little e ect on the inventory adjustment process.

5.2.2 Aggregate and Idiosyncratic Shocks
In this section, we examine the evolution of the aggregate and idiosyncratic shocks. Both the time path of aggregate shocks and the time path of the cross-sectional distribution of idiosyncratic shocks are computed from equation 2, using our estimates of the inventory deviation index. The time path of estimated aggregate shocks is computed according to

t = N ^1

Nt X

t i=1 

^it + hit,1 ; z

where Nt denotes the number of rms in period t. In each period t, the estimated crosssectional distribution of idiosyncratic shocks corresponds to the histogram of estimated it 's, where it = ^it + hit,1  , t : ^ z ^ We turn rst to aggregate shocks. The two panels of Figure 8 show the growth rate of aggregate inventories and the estimated time series of aggregate shocks for the durable and nondurable goods sectors. The aggregate shocks in both sectors are positively correlated with the sector-speci c aggregate inventory growth rate; the correlation between the two series is 0.66 for the durable goods industries and 0.42 for the nondurable goods industries. Note that aggregate inventory investment in both sectors was subject to a sequence of large negative aggregate shocks during the recessions of the 1980s, indicating that aggregate shocks are a major factor behind economy-wide inventory uctuations. Nevertheless, the correlation is far from perfect, suggesting that idiosyncratic shocks have a signi cant e ect on aggregate inventory movements. The two panels in Figure 9 show the time path of the standard deviation and skewness of the cross-sectional distribution of idiosyncratic shocks. Recall that, by de nition, the cross-sectional mean of idiosyncratic shocks is zero for all t. In both the durable and nondurable goods sectors, the idiosyncratic shocks exhibit substantial dispersion and skewness; moreover, both moments display considerable temporal variation. In particular, note that the skewness coe cient in both sectors was relatively large and negative during the economic turmoil of the late 1970s and early 1980s. This evidence indicates that the impact of the negative aggregate shocks on inventory investment during this period was augmented by a greater than average number of negative idiosyncratic shocks, a pattern consistent with the particularly acute inventory disinvestment during the 1981-82 recession; see Kashyap, Lamont, and Stein 1994 for a case study of the 1981-82 recession. 20

6 Conclusion
In this paper, we have examined inventory adjustment in the U.S. manufacturing sector, using a long panel of high-frequency rm-level data. Our theoretical framework is based on the generalized S; s methodology developed by Caballero, Engel, and Haltiwanger. The major advantage of our approach is that we are able to model explicitly the underlying microeconomic heterogeneity in the inventory adjustment process, while allowing for general nonconvexities in the production technology. A key result of our paper is that the estimated inventory adjustment function is nonlinear and asymmetric. The nonlinearity of the adjustment function is re ected in the fact that rms with larger absolute deviations from optimal inventory levels adjust proportionally more than do rms with smaller deviations. This nding implies rm-level inventory adjustment consistent with the use of generalized S; s type inventory policies, owing to the presence of nonconvexities in the production technology. The asymmetric shape of the inventory adjustment function implies that rms with small to moderately sized inventory overhangs adjust less than do rms with similarly sized inventory shortages. This asymmetry could re ect an additional stockout avoidance motive not captured by our model. Alternatively, the asymmetry, which is particularly pronounced in the durable goods sector, could be indicative of technological and or market irreversibilities in the inventory adjustment process. The inventory adjustment process appears to di er signi cantly between recessions and expansions. The cyclical shifts in the estimated inventory adjustment function indicate that rms with excess inventories disinvest more during recessions than they do during expansions. Such macroeconomic state-dependence in S; s inventory policies would be consistent with reductions in the inventory target levels, owing to increased demand uncertainty or adverse cost shocks. Outside of such cyclical shifts, the inventory adjustment process appears to have been relatively stable during our sample period. The estimated adjustment function during the 1980s is very similar to the adjustment function during the 1990s. In contrast, the average optimal inventory-sales ratio has declined signi cantly during our sample period, particularly in the durable goods sector. These results suggest that despite lower inventory target levels, the adoption of better inventory control methods have had little e ect on the adjustment process. Our conjecture is that recent inventory control innovations have a ected relative stockout costs but not the adjustment cost structure. However, additional research into the effects of inventory control innovations on both margins of rm inventory behavior is needed to resolve this dichotomy more fully. Such research may also provide more information 21

concerning the implications of such innovations for future aggregate inventory, and thus business cycle, uctuations.

References
Abernathy, F. H., J. T. Dunlop, J. H. Hammond, and D. Weil 1999: A Stitch

in Time: Lean Retailing and the Transformation of Manufacturing|Lessons From the Apparel and Textile Industries. Oxford University Press Inc., New York, NY.

Abramowitz, M. 1950: Inventories and Business Cycles. National Bureau of Economic

Research, New York, NY.

Model and Their Relation to E ciency," Biometric Journal, 23, 227 235. Allen, D. S. 1995: Changes in Inventory Management and the Business Cycle," Review, Federal Reserve Bank of St. Louis, 4, 17 33. Baltagi, B. H. 1995: Econometric Analysis of Panel Data. John Wiley & Sons Ltd., West Sussex, UK. Baltagi, B. H., and Y. J. Chang 1994: Incomplete Panels: A Comparative Study of Alternative Estimators for the Unbalanced One-Way Error Components Regression Model," Journal of Econometrics, 62, 67 89. Blanchard, O. J. 1983: The Production and Inventory Behavior of the American Automobile Industry," Journal of Political Economy, 91, 365 400. Blinder, A. S. 1981: Retail Inventory Behavior and Business Fluctuations," Brookings Papers on Economic Activity, 2, 443 505. 1986: Can the Production Smoothing Model of Inventories be Saved?," Quarterly Journal of Economics, 101, 431 454. Blinder, A. S., and L. J. Maccini 1991: Taking Stock: A Critical Assessment of Recent Research on Inventories," Journal of Economic Perspectives, 5, 73 96. Bresnahan, T., and V. Ramey 1994: Output Fluctuations at the Plant Level," Quarterly Journal of Economics, 109, 593 624. Caballero, R. J. 1999: Aggregate Investment," in Handbook of Macroeconomics, ed. by J. B. Taylor, and M. Woodford, pp. 813 862. North-Holland, Elsevier, New York, NY. Caballero, R. J., and E. M. R. A. Engel 1991: Dynamic S,s Economies," Econometrica, 59, 1659 1686. 1993: Microeconomic Adjustment Hazards and Aggregate Dynamics," Quarterly Journal of Economics, 108, 359 383. 22

Ahrens, H., and R. Pincus 1981: On Two Measures of Unbalancedness in a One-Way 

1999: Explaining Investment Dynamics in U.S. Manufacturing: A Generalized S; s Approach," Econometrica, 67, 783 826. Caballero, R. J., E. M. R. A. Engel, and J. C. Haltiwanger 1995: PlantLevel Adjustment and Aggregate Investment Dynamics," Brookings Papers on Economic Activity, 2, 1 54. 1997: Aggregate Employment Dynamics: Building from Microeconomic Evidence," American Economic Review, 87, 115 137. Caplin, A. S. 1985: The Variability of Aggregate Demand with S,s Inventory Policies," Econometrica, 53, 1395 1409. Diggle, P. J., K.-Y. Liang, and S. L. Zeger 1995: Analysis of Longitudinal Data. Clarendon Press, Oxford, UK. Episcopos, A. 1996: Testing the S; s Model of Inventory Investment with Canadian Wholesale Trade Data," Applied Economics Letters, 3, 193 195. Fair, R. C. 1989: Production Smoothing Model is Alive and Well," Journal of Monetary Economics, 24, 353 370. Filardo, A. J. 1995: Recent Evidence on the Muted Inventory Cycle," Federal Reserve Bank of Kansas City Economic Review, 2, 27 43. Hall, G. J. 1999: Non-Convex Costs and Capital Utilization: A Study of Production Scheduling at Automobile Assembly Plants," Forthcoming, Journal of Monetary Economics. Hall, G. J., and J. Rust 1999: An S; s Model of Commodity Price Speculation," Unpublished paper, Dept. of Economics, Yale University. Holt, C. C., F. Modigliani, J. F. Muth, and H. A. Simon 1960: Planning Production, Inventories and Work Force. Prentice-Hall, Englewood Cli s, NJ. Hordijk, A., and F. A. Van Der Duyn Schouten 1986: On the Optimality of s; S Policies in Continuous Review Inventory Models," SIAM Journal of Applied Mathematics, 46, 912 929. Hsiao, C. 1996: Random Coe cients Models," in The Econometrics of Panel Data: A Handbook of the Theory with Applications, ed. by L. Matyas, and P. Sevestre, pp. 77 99. Kluwer Academic Publishers, Norwell, MA. Kahn, J. A. 1987: Inventories and the Volatility of Production," American Economic Review, 77, 667 679. 1992: Why is Production More Volatile Than Sales? Theory and Evidence on the Stockout-Avoidance Motive for Inventory-Holding," Quarterly Journal of Economics, 152, 481 510. 23

Kashyap, A., O. Lamont, and J. C. Stein 1994: Credit Conditions and the Cyclical

Behavior of Inventories," Quarterly Journal of Economics, 109, 565 592. Krane, S. D. 1994: The Distinction Between Inventory Holding and Stockout Costs: Implications for Target Inventories, Asymmetric Adjustment, and the E ect of Aggregation on Production Smoothing," International Economic Review, 35, 117 136. Krane, S. D., and S. N. Braun 1991: Production Smoothing Evidence from Physical Product Data," Journal of Political Economy, 99, 558 581. Mairesse, J., and Z. Griliches 1990: Heterogeneity in Panel Data: Are There Stable Production Function?," in Essays in Honor of Edmond Malinvaud Empirical Economics, ed. by P. Champsaur, vol. 3, pp. 192 231. M.I.T. Press, Cambridge, MA. McCarthy, J., and E. Zakrajsek 1997: Trade Inventories," Unpublished paper, Federal Reserve Bank of New York. McConnell, M. M., and G. P. Quiros 1998: Output Fluctuations in the United States: What has Changed Since the Early 1980s?," Federal Reserve Bank of New York Sta Report No. 41. Mosser, P. C. 1988: Empirical Tests of the S,s Model for Merchant Wholesalers," in The Economics of Inventory Management, ed. by A. Chikan, and M. C. Lovell, pp. 261 283. Elsevier Science Publishers, Amsterdam, The Netherlands. 1991: Trade Inventories and S,s," Quarterly Journal of Economics, 106, 1267 1286. Ramey, V. A. 1991: Nonconvex Costs and the Behavior of Inventories," Journal of Political Economy, 99, 306 334. Ramey, V. A., and K. D. West 1999: Inventories," in Handbook of Macroeconomics, ed. by J. B. Taylor, and M. Woodford, pp. 863 923. North-Holland, Elsevier, New York, NY. Rao, C. R. 1971: Minimum Variance Quadratic Unbiased Estimation of Variance Components," Journal of Multivariate Analysis, 1, 445 456. Scarf, H. 1960: The Optimality of S,s Policies in the Dynamic Inventory Problem," in Mathematical Methods in the Social Sciences, ed. by K. J. Arrow, S. Karlin, and H. Scarf, pp. 196 202. Stanford University Press, Stanford, CA. Schuh, S. 1996: Evidence on the Link Between Firm-Level and Aggregate Inventory Behavior," Unpublished paper, Board of Governors of the Federal Reserve System. Swallow, W. H., and J. F. Monahan 1984: Monte Carlo Comparison of ANOVA, MIVQUE, REML, and ML Estimators of Variance Components," Technometrics, 26, 47 57. Veracierto, M. 1998: Plant-Level Irreversible Investment and Equilibrium Business Cycles," Federal Reserve Bank of Chicago Working Paper 98-1. 24

Industry
SIC 200 SIC 220 SIC 230 SIC 250 SIC 260 SIC 270 SIC 280 SIC 283 SIC 290 SIC 300 SIC 320
t

FGLS Estimates of the Random Coe cients Model Parameter

Table 1

e

2

WS a WC b Wtime c Wqtr d
0.01 0.01 0.01 0.01 0.01 0.02 0.01 0.01 0.01 0.01 0.18 0.01 0.01 0.01 0.01 0.02 0.06 0.01 0.01 0.10 0.01 0.01 0.14 0.27 0.13 0.06 0.56 0.12 0.29 0.26 0.10 0.03 0.02 0.01 0.02 0.03 0.01 0.01 0.04 0.13 0.25 0.01 0.19 0.01

!e

Obs:

0.788 0.065 0.768 0.068 0.819 0.097 0.925 0.093 0.772 0.086 0.868 0.142 0.866 0.055 0.730 0.049 0.604 0.085 0.775 0.077 0.435 0.099

-0.699 0.104 -0.914 0.095 -1.037 0.127 -1.242 0.182 -0.468 0.154 -0.225 0.176 -0.398 0.080 -0.349 0.077 -0.506 0.179 -0.887 0.125 -0.835 0.195

0.023 0.023 0.016 0.024 0.020 0.041 0.020 0.042 0.038 0.031 0.018

0.78 4,769 0.81 2,209 0.77 1,782 0.81 1,807 0.78 2,666 0.77 3,160 0.79 5,629 0.82 3,004 0.79 1,837 0.79 2,797 0.84 1,411

Notes: Estimation period: 1979:Q1 1997:Q4. Dependent variable is the log-level of real end of period inventories hit . All industry-speci c regressions include quarterly seasonal e ects, linear and quadratic time trends individual parameter estimates not reported and are estimated with FGLS, using MIVQUE0 Minimum Variance Quadratic Unbiased Estimator of the covariance matrix V . Heteroscedasticityconsistent asymptotic standard errors are reported in parenthesis. a Probability value for the Wald test of the null hypothesis that the coe cients on sit ; : : : ; sit,3 are jointly equal to zero. b Probability value for the Wald test of the null hypothesis that the coe cients on cit ; : : : ; cit,3 are jointly equal to zero. c Probability value for the Wald test of the null hypothesis that the linear and quadratic time trends are jointly equal to zero. d Probability value for the Wald test of the null hypothesis that the quarterly seasonal e ects are jointly equal to zero. P e A measure of panel unbalancedness given by Ahrens and Pincus 1981: ! = N=T  T1i , where P T = Ti =N. Note that 0  !  1, with ! = 1 when the panel is balanced.

25

Industry
SIC 330 SIC 340 SIC 350 SIC 357 SIC 360 SIC 366 SIC 367 SIC 370 SIC 380 SIC 382 SIC 384 SIC 390
t

FGLS Estimates of the Random Coe cients Model Parameter

Table 1 continued
e

2

WS a WC b Wtime c Wqtr d
0.01 0.01 0.01 0.01 0.01 0.02 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.01 0.02 0.06 0.01 0.01 0.01 0.01 0.01 0.01 0.29 0.13 0.24 0.39 0.15 0.83 0.09 0.03 0.41 0.33 0.83 0.35 0.01 0.59 0.01 0.01 0.01 0.09 0.58 0.01 0.17 0.60 0.15 0.02

!e

Obs:

0.605 0.081 0.770 0.055 0.736 0.034 0.804 0.046 0.820 0.053 0.690 0.066 0.621 0.049 0.729 0.064 0.652 0.080 0.714 0.049 0.695 0.054 0.683 0.056

-0.803 0.130 -1.258 0.121 -0.811 0.075 -0.525 0.076 -0.826 0.076 -0.889 0.123 -0.568 0.089 -0.929 0.133 -0.936 0.116 -0.584 0.072 -0.452 0.081 -1.081 0.106

0.026 0.011 0.029 0.059 0.022 0.038 0.036 0.031 0.026 0.027 0.049 0.032

0.76 3,289 0.78 4,204 0.81 7,402 0.81 4,361 0.83 4,444 0.81 3,379 0.80 4,677 0.79 4,259 0.81 1,673 0.81 4,973 0.81 3,628 0.79 3,958

Notes: Estimation period: 1979:Q1 1997:Q4. Dependent variable is the log-level of real end of period inventories hit . All industry-speci c regressions include quarterly seasonal e ects, linear and quadratic time trends individual parameter estimates not reported and are estimated with FGLS, using MIVQUE0 Minimum Variance Quadratic Unbiased Estimator of the covariance matrix V . Heteroscedasticityconsistent asymptotic standard errors are reported in parenthesis. a Probability value for the Wald test of the null hypothesis that the coe cients on sit ; : : : ; sit,3 are jointly equal to zero. b Probability value for the Wald test of the null hypothesis that the coe cients on cit ; : : : ; cit,3 are jointly equal to zero. c Probability value for the Wald test of the null hypothesis that the linear and quadratic time trends are jointly equal to zero. d Probability value for the Wald test of the null hypothesis that the quarterly seasonal e ects are jointly equal to zero. P e A measure of panel unbalancedness given by Ahrens and Pincus 1981: ! = N=T  T1i , where P T = Ti =N. Note that 0  !  1, with ! = 1 when the panel is balanced.

26

Figure 1: Average Optimal Inventory-Sales Ratio
0.85 0.58

Durables (left scale) Nondurables (right scale)
0.80

0.56

0.54 0.75 0.52 0.70 0.50 0.65 0.48

0.60 1979 1981 1983 1985 1987 1989 1991 1993 1995 1997

0.46

Seasonally adjusted data. Shaded regions indicate NBER-dated recessions.

27

Figure 2: Inventory Adjustment Function
0.88 0.86 0.84 0.82 0.80 0.78 0.025 0.030 0.035

0.76 0.74 0.72 0.70 0.68 0.66 0.64 0.62 0.60 0.58 0.56 0.0 -0.40 -0.35 -0.30 -0.25 -0.20 -0.15 -0.10 -0.05 0.0 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.005 0.010 0.015 0.020

Inventory Deviation Index (z)

Figure 3: Sector-Speci c Inventory Adjustment Functions
0.86 0.84 0.82 0.80 0.78 0.86 0.84 0.82 0.80 0.78 0.76 0.74 0.72 0.70 0.68 0.66 0.64

Adjustment Rate

0.76 0.74 0.72 0.70 0.68 0.66 0.64 0.62 0.60 0.58

Durables Nondurables

0.62 0.60 0.58

-0.40

-0.35

-0.30

-0.25

-0.20

-0.15

-0.10

-0.05

0.0

0.05

0.10

0.15

0.20

0.25

0.30

0.35

0.40

Inventory Deviation Index (z)

28

Average Density of z

Adjustment Rate

Figure 4: Cyclical Shifts in the Adjustment Function
0.90 0.88 0.86 0.84 0.82 0.80 0.90 0.88 0.86 0.84 0.82 0.80 0.78 0.76 0.74 0.72 0.70 0.68 0.66

Adjustment Rate

0.78 0.76 0.74 0.72 0.70 0.68 0.66 0.64 0.62 0.60 0.58 -0.40 -0.35 -0.30 -0.25 -0.20 -0.15 -0.10 -0.05 0.0 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40

Expansion Recession

0.64 0.62 0.60 0.58

Inventory Deviation Index (z)

Figure 5: Decade-Speci c Inventory Adjustment Functions
0.84 0.82 0.80 0.78 0.76 0.84 0.82 0.80 0.78 0.76 0.74 0.72 0.70 0.68 0.66 0.64

Adjustment Rate

0.74 0.72 0.70 0.68 0.66 0.64 0.62 0.60 0.58 0.56 -0.40 -0.35 -0.30 -0.25 -0.20 -0.15 -0.10 -0.05 0.0 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40

1980s 1990s

0.62 0.60 0.58 0.56

Inventory Deviation Index (z)

29

Figure 6: Aggregate Inventory Growth and Inventory Deviation Index
Manufacturing Durable Goods
Percent 10

5

0

-5

-10

Inventory Growth Inv. Deviation Index

-15 1979 1981 1983 1985 1987 1989 1991 1993 1995 1997

Manufacturing Nondurable Goods
Percent 10

5

0

-5

-10

Inventory Growth Inv. Deviation Index

-15 1979 1981 1983 1985 1987 1989 1991 1993 1995 1997

Seasonally adjusted data. Shaded regions indicate NBER-dated recessions.

30

Figure 7: Cross-Sectional Moments of Inventory Deviations
Cross-Sectional Standard Deviation of z
0.20 0.18 0.16 0.14 0.12 0.10 0.08 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997

Durables Nondurables

Cross-Sectional Skewness of z
4 3 2 1 0 -1 -2 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997

Durables Nondurables

Cross-Sectional Excess Kurtosis of z
25 20 15 10 5 0 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 Seasonally adjusted data. Shaded regions indicate NBER-dated recessions.

Durables Nondurables

31

Figure 8: Aggregate Inventory Growth and Aggregate Shocks
Manufacturing Durable Goods
Percent 10

5

0

-5

-10

Inventory Growth Aggregate Shock

-15

1979

1981

1983

1985

1987

1989

1991

1993

1995

1997

Manufacturing Nondurable Goods
Percent 10

5

0

-5

-10

Inventory Growth Aggregate Shock

-15 1979 1981 1983 1985 1987 1989 1991 1993 1995 1997

Seasonally adjusted data. Shaded regions indicate NBER-dated recessions.

32

Figure 9: Cross-Sectional Moments of Idiosyncratic Shocks
Standard Deviation of Idiosyncratic Shocks
0.16

0.15

0.14

0.13

0.12

0.11

Durables Nondurables
1979 1981 1983 1985 1987 1989 1991 1993 1995 1997

Skewness of Idiosyncratic Shocks
1.5 1.0 0.5 0.0 -0.5 -1.0 -1.5 1979 1981 1983 1985 1987 1989 1991 1993 1995 1997

Durables Nondurables

Seasonally adjusted data. Shaded regions indicate NBER-dated recessions.

33

A Data Appendix
This section describes the selection rules used to construct our rm-level panel and the construction of the variables used in the analysis. The data for our paper come from the quarterly P S T, Full Coverage, and Research COMPUSTAT data les. The rm-level COMPUSTAT data are compiled in a scal-year format. The scal quarters in the data are aligned with calendar quarters as follows: 1. If the rm's scal year ends in the same month as a calendar quarter, the adjustment is straightforward, as the scal quarters are relabeled to correspond to calendar quarters. 2. If the rm's scal-year end does not coincide with the end of a calendar quarter, the data are adjusted so that the majority of the scal quarter is placed into the appropriate calendar quarter. We selected all rms with positive total inventories, positive net sales, positive total assets, and with at least 20 continuous quarters of data between 1978Q1 and 1997Q4. To avoid results that are driven by a small number of extreme observations, three criteria were used to eliminate rms with substantial outliers or obvious errors: 1. If a rm's estimate of real gross output from the accounting identity Y S + H was negative at any point during a rm's tenure in the sample, a rm was eliminated in its entirety. 2. If a rm's quarterly growth rate of real inventories was above below the 99.5 0.5 percentile of the distribution in any period during the rm's tenure in the panel, the rm was eliminated in its entirety. 3. If a rm's quarterly growth rate of real sales was above below the 99.5 0.5 percentile of the distribution in any period during the rm's tenure in the panel, the rm was eliminated in its entirety. As a consequence of these selection rules, 911 rms were eliminated from the original panel.25 Table A.1 provides a detailed industry breakdown of our panel. All of our industry groups contain more than 1,500 rm quarter observations. The sparsest industry in our data set is SIC 290 Petroleum & Coal Products, which contains only 40 rms. On the other hand, SIC 350 Industrial Machinery & Equipment contains 189 rms. Finally, as mentioned in the text, the panels are unbalanced, with rms entering and exiting the data set. The lowest average industry-speci c tenure in the panel is almost 38 quarters SIC 357: Computers & O ce Equipment, and the highest is almost 51 quarters SIC 290: Petroleum & Coal Products.
Over 3 4 of eliminated rms were deleted because of the second and third selection criteria. Visual inspection of the eliminated rms revealed severe anomalities and likely errors in their reported data.
25

A.1 Selection Rules

34

A.2 Construction of Variables

cause the rm-level COMPUSTAT data provide limited and incomplete information on the inventory accounting practices, we assumed that all inventory stocks are evaluated using the FIFO method; namely, once a nished good is placed on shelves, it is given a price tag that remains on the item regardless of what subsequently happens to the price of newly produced goods. This implies that the replacement value of inventory stocks equals their book value. To convert the reported nominal value of inventories to real terms, inventory stocks were de ated by the sector-speci c i.e., durable and nondurable implicit 1992=100 inventory de ator. The inventory stocks are measured as of the end of the period. Net Sales: To construct a real measure of sales, the reported nominal value of sales was de ated by the sector-speci c i.e., durable and nondurable implicit 1992=100 sales de ator. Gross Output: An estimate of a rm's real gross output in period t, Yt, was obtained from the accounting identity Yt St +Ht, where St denotes real nal sales in period t, and Ht denotes real inventory investment in period t.26 Cost per Unit of Output: A real cost per unit of output was constructed by converting the nominal Cost of Goods Sold to real terms using the implicit 1992=100 GDP de ator. The ratio of real total costs to real gross output is our measure of the average cost per unit of output. All other variables were de ated by the implicit 1992=100 GDP de ator. Table A.2 provides summary statistics for the key variables used in our analysis. Because all rms in the sample are publicly traded, most of them are relatively large. The median rm size, measured by total assets, is $125 million. The distributions of most variables display considerable skewness|the means of inventories, sales, gross output, and assets are much greater than the medians. The distribution of the inventory-sales ratio, on the other hand, is considerably more symmetric. Also note that even after excluding outliers there remains a great deal of heterogeneity in inventory investment and in the growth of inventories and sales.

Inventories: The COMPUSTAT data report the book value of total inventories. Be-

Note that the accounting identity, Y S + H, holds only for nished goods inventories. Because the quarterly COMPUSTAT data report only the dollar value of total inventory stocks, this relationship does not hold exactly in our data. Nevertheless, the results reported in this paper are virtually identical when we include rms that violate this accounting identity."
26

35

Industry Composition

Table A.1

Industry Classi cation
SIC 200: SIC 220: SIC 230: SIC 250: SIC 260: SIC 270: SIC 280: SIC 283: SIC 290: SIC 300: SIC 320: SIC 330: SIC 340: SIC 350: SIC 357: SIC 360: SIC 366: SIC 367: SIC 370: SIC 380: SIC 382: SIC 384: SIC 390:

Food & Kindred Prod.a Textile Mill Prod. Apparel & Other Prod. Furniture & Fixtures Paper & Allied Prod. Printing & Publishing Chemical & Allied Prod.b Drugs Petroleum & Coal Prod. Rubber & Misc. Plastic Prod. Stone, Clay & Glass Prod. Primary Metal Industries Fabricated Metal Prod. Industrial Machinery & Equip.c Computers & O ce Equip. Electronic & Other Electric Equip.d Communications Equip Electronic Components Transportation Equip. Instruments & Related Prod.e Measuring & Controlling Prod. Medical Instruments Misc. Manufacturing Industriesf 

of Firms Avg: Ti Med: Ti 127 42.6 38.0 67 38.0 30.0 52 39.3 30.0 48 42.7 44.0 62 48.0 44.5 75 47.1 40.0 128 49.0 46.0 81 42.1 36.0 40 50.9 43.5 81 39.5 34.0 43 37.8 34.0 84 44.2 37.5 113 42.2 38.0 189 44.2 39.0 134 37.5 34.0 121 41.7 39.0 95 50.6 37.0 125 42.4 38.0 112 43.0 37.0 50 38.5 33.5 131 43.0 39.0 108 38.6 34.5 103 43.4 40.0

Obs. 5,404 2,544 2,042 2,047 2,976 3,535 6,269 3,409 2,037 3,202 1,626 3,709 4,769 8,347 5,031 5,049 3,854 5,302 4,819 1,923 5,628 4,168 4,473

Includes SIC 21 Tobacco and Tobacco Products. Excludes SIC 283 Drugs. c Excludes SIC 357 Computers and O ce Equipment. d Excludes SIC 366 Communications Equipment and SIC 367 Electronic Components. e Excludes SIC 382 Measuring and Controlling Products and SIC 384 Medical Instruments. f Includes SIC 24 Lumber & Wood Products and SIC 31 Leather & Leather Products.
a b

36

Variable
Inventories Net Sales Gross Outputa Total Assets Inv. Investment Inv Sales Ratio Average Costb

Summary Statistics for All Industries Mean Std.Dev. Median Minimum Maximum 191.4 403.5 408.0 1,588.5 0.84 0.73 0.73 1.31 1.49 2,169 92,163 661.5 1699.7 1,708.9 7,955.1 58.8 0.46 0.44 13.3 17.9 24.5 39.6 40.4 125.0 0.09 0.64 0.72 1.04 1.80 0.039 0.012 0.010 0.235 -2,444.7 0.01 0.25 -72.1 -112.3 12,660.6 39,335.0 38,280.7 259,303.0 3,077.6 15.4 49.4 82.5 114.5

Table A.2

Inv. Growth Rate  Sales Growth Rate   of Firms Observations
a

Real gross output in period t, Yt, was estimated from the accounting identity Yt St + Ht, where St denotes real nal sales in period t, and Ht denotes real inventory investment in period t. b Real average cost is de ned as the ratio of real cost of goods sold to real gross output.

Notes: Sample period: 1978:Q1 1997:Q4. All variables are in millions of 1992 dollars.

37